In the process of coal gangue sorting, traditional methods use visual identification and belt speed to predict the real-time position of coal gangue. However, due to slippage and deviation of the coal gangue during high-speed, long-distance transportation on the belt, the actual position often differs from the predicted position, leading to issues like missed or incorrect grabs by robotic arms, thus affecting sorting efficiency. To address this problem, this study proposes an improved ORB matching algorithm for the secondary positioning of coal gangue. Firstly, it introduces local adaptive gamma correction to oFAST feature detection, enhancing matching accuracy under low lighting conditions. Additionally, to counter the dynamic interference caused by high-speed movement of the gangue, this paper combines BEBLID descriptors and the GMS algorithm for rapid feature matching, and employs the RANSAC algorithm to optimize feature point selection, thereby enhancing the robustness of the algorithm. Ultimately, the minimum bounding rectangle is calculated through matching points to obtain the position. Experimental results show that the proposed algorithm improves the matching accuracy of coal gangue under scale, illumination, and angle changes by 16.7%, 36%, and 22% respectively, compared to the traditional ORB algorithm, with an average error of 1.29mm and an average matching time within 40ms. This effectively enables gangue matching and positioning in complex scenarios.